过度参数化的神经网络的实际成功促进了最近对插值方法的科学研究，这些研究非常适合其训练数据。如果没有灾难性的测试表现，包括神经网络在内的某些插值方法（包括神经网络）可以符合嘈杂的训练数据，这是违反统计学习理论的标准直觉的。为了解释这一点，最近的一系列工作研究了$ \ textit {良性过拟合} $，这是一种现象，其中一些插值方法即使在存在噪音的情况下也接近了贝叶斯的最佳性。在这项工作中，我们认为，虽然良性过度拟合既具有启发性和富有成效的研究在测试时间的风险，这意味着这些模型既不是良性也不是灾难性的，而是属于中间状态。我们称此中级制度$ \ textit {perked forporting} $，我们启动其系统研究。我们首先在内核（Ridge）回归（KR）的背景下探索这种现象，通过在脊参数和核特征光谱上获得条件，KR在这些条件下表现出三种行为。我们发现，具有PowerLaw光谱的内核，包括Laplace内核和Relu神经切线内核，表现出了过度拟合的。然后，我们通过分类法的镜头从经验上研究深度神经网络，并发现接受插值训练的人是脾气暴躁的，而那些训练的人则是良性的。我们希望我们的工作能够使人们对现代学习过度拟合的过度理解。

Kernels are efficient in representing nonlocal dependence and they are widely used to design operators between function spaces. Thus, learning kernels in operators from data is an inverse problem of general interest. Due to the nonlocal dependence, the inverse problem can be severely ill-posed with a data-dependent singular inversion operator. The Bayesian approach overcomes the ill-posedness through a non-degenerate prior. However, a fixed non-degenerate prior leads to a divergent posterior mean when the observation noise becomes small, if the data induces a perturbation in the eigenspace of zero eigenvalues of the inversion operator. We introduce a data-adaptive prior to achieve a stable posterior whose mean always has a small noise limit. The data-adaptive prior's covariance is the inversion operator with a hyper-parameter selected adaptive to data by the L-curve method. Furthermore, we provide a detailed analysis on the computational practice of the data-adaptive prior, and demonstrate it on Toeplitz matrices and integral operators. Numerical tests show that a fixed prior can lead to a divergent posterior mean in the presence of any of the four types of errors: discretization error, model error, partial observation and wrong noise assumption. In contrast, the data-adaptive prior always attains posterior means with small noise limits.

With more and more data being collected, data-driven modeling methods have been gaining in popularity in recent years. While physically sound, classical gray-box models are often cumbersome to identify and scale, and their accuracy might be hindered by their limited expressiveness. On the other hand, classical black-box methods, typically relying on Neural Networks (NNs) nowadays, often achieve impressive performance, even at scale, by deriving statistical patterns from data. However, they remain completely oblivious to the underlying physical laws, which may lead to potentially catastrophic failures if decisions for real-world physical systems are based on them. Physically Consistent Neural Networks (PCNNs) were recently developed to address these aforementioned issues, ensuring physical consistency while still leveraging NNs to attain state-of-the-art accuracy. In this work, we scale PCNNs to model building temperature dynamics and propose a thorough comparison with classical gray-box and black-box methods. More precisely, we design three distinct PCNN extensions, thereby exemplifying the modularity and flexibility of the architecture, and formally prove their physical consistency. In the presented case study, PCNNs are shown to achieve state-of-the-art accuracy, even outperforming classical NN-based models despite their constrained structure. Our investigations furthermore provide a clear illustration of NNs achieving seemingly good performance while remaining completely physics-agnostic, which can be misleading in practice. While this performance comes at the cost of computational complexity, PCNNs on the other hand show accuracy improvements of 17-35% compared to all other physically consistent methods, paving the way for scalable physically consistent models with state-of-the-art performance.

Sign language is the preferred method of communication of deaf or mute people, but similar to any language, it is difficult to learn and represents a significant barrier for those who are hard of hearing or unable to speak. A person's entire frontal appearance dictates and conveys specific meaning. However, this frontal appearance can be quantified as a temporal sequence of human body pose, leading to Sign Language Recognition through the learning of spatiotemporal dynamics of skeleton keypoints. I propose a novel, attention-based approach to Sign Language Recognition exclusively built upon decoupled graph and temporal self-attention: the Sign Language Graph Time Transformer (SLGTformer). SLGTformer first deconstructs spatiotemporal pose sequences separately into spatial graphs and temporal windows. SLGTformer then leverages novel Learnable Graph Relative Positional Encodings (LGRPE) to guide spatial self-attention with the graph neighborhood context of the human skeleton. By modeling the temporal dimension as intra- and inter-window dynamics, I introduce Temporal Twin Self-Attention (TTSA) as the combination of locally-grouped temporal attention (LTA) and global sub-sampled temporal attention (GSTA). I demonstrate the effectiveness of SLGTformer on the World-Level American Sign Language (WLASL) dataset, achieving state-of-the-art performance with an ensemble-free approach on the keypoint modality.

The number of international benchmarking competitions is steadily increasing in various fields of machine learning (ML) research and practice. So far, however, little is known about the common practice as well as bottlenecks faced by the community in tackling the research questions posed. To shed light on the status quo of algorithm development in the specific field of biomedical imaging analysis, we designed an international survey that was issued to all participants of challenges conducted in conjunction with the IEEE ISBI 2021 and MICCAI 2021 conferences (80 competitions in total). The survey covered participants' expertise and working environments, their chosen strategies, as well as algorithm characteristics. A median of 72% challenge participants took part in the survey. According to our results, knowledge exchange was the primary incentive (70%) for participation, while the reception of prize money played only a minor role (16%). While a median of 80 working hours was spent on method development, a large portion of participants stated that they did not have enough time for method development (32%). 25% perceived the infrastructure to be a bottleneck. Overall, 94% of all solutions were deep learning-based. Of these, 84% were based on standard architectures. 43% of the respondents reported that the data samples (e.g., images) were too large to be processed at once. This was most commonly addressed by patch-based training (69%), downsampling (37%), and solving 3D analysis tasks as a series of 2D tasks. K-fold cross-validation on the training set was performed by only 37% of the participants and only 50% of the participants performed ensembling based on multiple identical models (61%) or heterogeneous models (39%). 48% of the respondents applied postprocessing steps.

Multimodal models are becoming increasingly effective, in part due to unified components, such as the Transformer architecture. However, multimodal models still often consist of many task- and modality-specific pieces and training procedures. For example, CLIP (Radford et al., 2021) trains independent text and image towers via a contrastive loss. We explore an additional unification: the use of a pure pixel-based model to perform image, text, and multimodal tasks. Our model is trained with contrastive loss alone, so we call it CLIP-Pixels Only (CLIPPO). CLIPPO uses a single encoder that processes both regular images and text rendered as images. CLIPPO performs image-based tasks such as retrieval and zero-shot image classification almost as well as CLIP, with half the number of parameters and no text-specific tower or embedding. When trained jointly via image-text contrastive learning and next-sentence contrastive learning, CLIPPO can perform well on natural language understanding tasks, without any word-level loss (language modelling or masked language modelling), outperforming pixel-based prior work. Surprisingly, CLIPPO can obtain good accuracy in visual question answering, simply by rendering the question and image together. Finally, we exploit the fact that CLIPPO does not require a tokenizer to show that it can achieve strong performance on multilingual multimodal retrieval without

Federated learning (FL) is an emerging machine learning paradigm, in which clients jointly learn a model with the help of a cloud server. A fundamental challenge of FL is that the clients are often heterogeneous, e.g., they have different computing powers, and thus the clients may send model updates to the server with substantially different delays. Asynchronous FL aims to address this challenge by enabling the server to update the model once any client's model update reaches it without waiting for other clients' model updates. However, like synchronous FL, asynchronous FL is also vulnerable to poisoning attacks, in which malicious clients manipulate the model via poisoning their local data and/or model updates sent to the server. Byzantine-robust FL aims to defend against poisoning attacks. In particular, Byzantine-robust FL can learn an accurate model even if some clients are malicious and have Byzantine behaviors. However, most existing studies on Byzantine-robust FL focused on synchronous FL, leaving asynchronous FL largely unexplored. In this work, we bridge this gap by proposing AFLGuard, a Byzantine-robust asynchronous FL method. We show that, both theoretically and empirically, AFLGuard is robust against various existing and adaptive poisoning attacks (both untargeted and targeted). Moreover, AFLGuard outperforms existing Byzantine-robust asynchronous FL methods.

Training large, deep neural networks to convergence can be prohibitively expensive. As a result, often only a small selection of popular, dense models are reused across different contexts and tasks. Increasingly, sparsely activated models, which seek to decouple model size from computation costs, are becoming an attractive alternative to dense models. Although more efficient in terms of quality and computation cost, sparse models remain data-hungry and costly to train from scratch in the large scale regime. In this work, we propose sparse upcycling -- a simple way to reuse sunk training costs by initializing a sparsely activated Mixture-of-Experts model from a dense checkpoint. We show that sparsely upcycled T5 Base, Large, and XL language models and Vision Transformer Base and Large models, respectively, significantly outperform their dense counterparts on SuperGLUE and ImageNet, using only ~50% of the initial dense pretraining sunk cost. The upcycled models also outperform sparse models trained from scratch on 100% of the initial dense pretraining computation budget.

Classifiers in supervised learning have various security and privacy issues, e.g., 1) data poisoning attacks, backdoor attacks, and adversarial examples on the security side as well as 2) inference attacks and the right to be forgotten for the training data on the privacy side. Various secure and privacy-preserving supervised learning algorithms with formal guarantees have been proposed to address these issues. However, they suffer from various limitations such as accuracy loss, small certified security guarantees, and/or inefficiency. Self-supervised learning is an emerging technique to pre-train encoders using unlabeled data. Given a pre-trained encoder as a feature extractor, supervised learning can train a simple yet accurate classifier using a small amount of labeled training data. In this work, we perform the first systematic, principled measurement study to understand whether and when a pre-trained encoder can address the limitations of secure or privacy-preserving supervised learning algorithms. Our key findings are that a pre-trained encoder substantially improves 1) both accuracy under no attacks and certified security guarantees against data poisoning and backdoor attacks of state-of-the-art secure learning algorithms (i.e., bagging and KNN), 2) certified security guarantees of randomized smoothing against adversarial examples without sacrificing its accuracy under no attacks, 3) accuracy of differentially private classifiers, and 4) accuracy and/or efficiency of exact machine unlearning.

Pixel-level labels are particularly expensive to acquire. Hence, pretraining is a critical step to improve models on a task like semantic segmentation. However, prominent algorithms for pretraining neural networks use image-level objectives, e.g. image classification, image-text alignment a la CLIP, or self-supervised contrastive learning. These objectives do not model spatial information, which might be suboptimal when finetuning on downstream tasks with spatial reasoning. In this work, we propose to pretrain networks for semantic segmentation by predicting the relative location of image parts. We formulate this task as a classification problem where each patch in a query view has to predict its position relatively to another reference view. We control the difficulty of the task by masking a subset of the reference patch features visible to those of the query. Our experiments show that this location-aware (LOCA) self-supervised pretraining leads to representations that transfer competitively to several challenging semantic segmentation benchmarks.